An AI agent can prepare a payment. That does not mean it should approve one
Automation can extract invoices, match purchase orders, and propose a transaction. Final authority should remain in a separate, inspectable control plane.
5 min read
The archive
All stories
Sourced guides, explanations and news analysis. Read by question, not by jargon.
Why two cybercrime reports can tell different stories without contradicting each other
Complaint databases, incident-response cases, product telemetry, and surveys observe different populations. Compare definitions and denominators before comparing percentages.
5 min read
Followers, testimonials, and busy group chats: how confidence gets borrowed
Visible approval can be purchased, coordinated, selective, or unrelated to the exact offer. Trace claims back to independent experience and verifiable outcomes.
5 min read
One more step to recover everything: the trap inside a nearly finished process
After time, money, or identity documents have been invested, the next demand can feel like the cheapest route out. Reframe the decision around future risk.
5 min read
The boss wants an exception: designing a process that can survive authority
A process is not reliable if the people with the most organizational status can silently bypass it. Build escalation and verification into the rule itself.
6 min read
When a search answer turns an unverified phone number into apparent authority
Search results can compress a source, a summary, and a phone number into one confident answer. During a problem, that convenience can erase the provenance you need most.
6 min read
The email passed authentication. The payment request still needs authorization
DMARC can validate authorized use of a sending domain. It does not establish that the message is truthful, expected, or approved under the recipient’s payment process.
6 min read
The login window that is only a picture of a window
A convincing sign-in pop-up can be drawn inside a webpage. The important boundary is the real browser chrome and the domain that receives authentication.
5 min read
A screenshot is not the whole scam: what an AI check cannot establish
An AI can describe a screenshot and surface questions, but the crop cannot prove who controls the account, where a link leads, or whether the surrounding transaction is genuine.
6 min read
Prompt injection borrows authority from the page an AI assistant is reading
Indirect prompt injection works by making untrusted content sound like an instruction to the assistant reading it. The durable control is to limit what that assistant can authorize.
6 min read
A lower phishing click rate can still tell the wrong story
A practical reading of phishing-simulation metrics: scenario difficulty, exposure, reporting and workflow evidence matter alongside the percentage.
7 min read
Knowing your first school is not the same as being you
Why biographical security questions are weak evidence, how NIST frames the issue and what a usable account-recovery process should establish instead.
8 min read